Memory supercycle

A "memory supercycle" is a prolonged, demand-driven boom in memory chips (DRAM, NAND and especially HBM) where prices and revenue surge for years rather than a few quarters. The 2025-2026 cycle is fueled by AI data centers, which now consume roughly 70% of memory output and have pushed contract prices up 50-75% in a single quarter.

What is a memory supercycle?

A **memory supercycle** is an unusually long and intense upturn in the memory-chip market, where demand outruns supply for years rather than the typical few quarters. Memory (DRAM, NAND flash and high-bandwidth memory) is a commodity that historically swings through sharp boom-and-bust cycles: prices spike, makers add capacity, oversupply crashes prices, then the cycle repeats. A *supercycle* breaks that rhythm because a new, structural source of demand absorbs supply faster than factories can expand. Bank of America has compared the 2026 upturn to the memory boom of the 1990s, forecasting DRAM revenue up roughly 51% and NAND up 45% year over year.

How memory chips work (the three types)

Three memory products drive the cycle. **DRAM** is fast, volatile working memory that holds data a processor is actively using. **NAND flash** is slower, non-volatile storage (SSDs) that keeps data when power is off. **HBM (high-bandwidth memory)** is the star of the AI era: it stacks DRAM dies vertically (8-Hi, 12-Hi, now 16-Hi) and places them next to a GPU, connected by a very wide interface so data moves at terabytes per second. HBM solves the *memory wall* — the bottleneck where a fast AI chip sits idle waiting for data. Samsung's first HBM4E samples reportedly hit 3.6 TB/s per stack.

Why it matters for AI and data centers

Training and running large AI models is memory-bound as much as compute-bound: each GPU needs enormous, fast memory to hold model weights and activations. As a result, AI data centers now consume an estimated **70% of all memory chips produced worldwide**. Every AI accelerator (such as NVIDIA's GPUs) ships with multiple HBM stacks, and servers also pack growing amounts of conventional DRAM and NAND. Micron projects the HBM market alone could grow at roughly a 40% CAGR toward **$100 billion**, exceeding the size of the entire 2024 DRAM market. This concentrated demand is what turns an ordinary up-cycle into a supercycle.

Where memory sits in the chip supply chain

HBM is part of *advanced packaging*: bare DRAM dies are stacked using through-silicon vias (TSVs), then mounted on a silicon interposer beside the GPU and bonded into one package (e.g. via TSMC's CoWoS). So the supply chain runs memory maker → packaging/foundry → accelerator vendor → server OEM → cloud operator. A key feature of the current cycle is that HBM is *capacity-hungry*: it consumes far more wafers per usable bit than standard DRAM, so as HBM scales it eats into the wafers available for ordinary DRAM — tightening the whole market and lifting PC, phone and SSD prices too.

Who the key players are

Memory is an oligopoly of three: **SK Hynix**, **Samsung** and **Micron (MU)**. SK Hynix leads HBM with roughly 57% of HBM revenue in Q3 2025, Samsung around 22% and Micron around 21%; all three are racing to supply NVIDIA's next-gen platforms with HBM4. In NAND/storage, **SanDisk (SNDK)** — spun out of Western Digital — is a major flash supplier benefiting from surging SSD demand. Because SK Hynix and Samsung are South Korean, broad Korea exposure (e.g. the **EWY** ETF) is a common indirect way investors play the theme. Note: example tickers are context, not recommendations.

What is changing right now

The cycle is intensifying. In Q2 2026, DRAM contract prices reportedly rose 58-63% and NAND 70-75% quarter over quarter — among the sharpest jumps in over a decade — and major makers say their 2026 HBM output is *already sold out*. The industry is transitioning from HBM3E to **HBM4** (mass production ramping in 2026) and toward 16-Hi stacks for NVIDIA's Rubin generation. Several analysts argue the AI-driven shortage could persist toward 2030. The flip side: consumers and PC/phone OEMs face higher prices, and the deeply cyclical history of memory means a downturn will eventually follow.

Frequently asked

Is the memory supercycle the same as a normal memory up-cycle?

No. Memory always cycles up and down, but a supercycle is longer and steeper because a durable, structural demand source — in this case AI data centers — keeps supply tight for years instead of quarters. Analysts have likened the 2026 boom to the 1990s, with DRAM and NAND revenue projected to rise 45-51% in a single year.

Why are AI chips driving memory demand so hard?

AI accelerators are limited by the 'memory wall': the GPU can compute faster than memory can feed it data. Vendors solve this with HBM — stacked DRAM placed next to the GPU delivering terabytes per second. Each accelerator needs several HBM stacks plus large amounts of system DRAM, so AI buildout consumes memory disproportionately, around 70% of global output.

What is HBM and how is it different from regular DRAM?

HBM (high-bandwidth memory) is made of standard DRAM dies stacked vertically and wired with a very wide interface, then packaged right beside the processor. It delivers far higher bandwidth than ordinary DRAM modules but costs more and uses more wafers per bit, which is part of why HBM growth tightens the entire memory market.

Which companies benefit from the memory supercycle?

The three memory makers — SK Hynix, Samsung and Micron (MU) — benefit most, with SK Hynix leading in HBM. NAND/storage suppliers like SanDisk (SNDK) gain from surging SSD prices. Because two leaders are Korean, broad Korea ETFs such as EWY offer indirect exposure. Packaging foundries and AI accelerator vendors are also tied to the cycle.

When will the memory supercycle end?

No one knows precisely. Several analysts expect AI-driven tightness to persist toward 2030, but memory is historically very cyclical — heavy capacity investment plus any slowdown in AI spending could trigger oversupply and falling prices. Treat long-cycle forecasts as scenarios, not certainties.

Related companies